Everyone assumed AI would come for copyediting first. Catch the typos, fix the comma splices, flag the passive voice. Mechanical work. Pattern matching.
That’s not what happened.
The Three Jobs of Editing
Editing has always been three distinct cognitive tasks bundled into one role.
Line editing operates at the sentence level. Word choice, rhythm, clarity. Does this sentence say what it means? Does it sound right when read aloud?
Developmental editing works at the structural level. Does the argument hold together? Is the evidence in the right order? Does the piece know what it’s about?
Editorial judgment sits above both. Should this piece exist at all? Is it saying something true and non-obvious? Does it matter to the audience we’re trying to reach?
The assumption was that AI would climb this ladder from the bottom. Start with grammar, eventually get to structure, maybe someday develop taste.
The ladder fell over sideways.
What Actually Happened
GPT-4 and Claude can now do competent developmental editing. Give them a messy draft and they’ll identify structural problems with surprising accuracy: “Your third section is doing the work your introduction should do.” “You’ve buried your actual thesis in paragraph six.” “These two arguments contradict each other—which one do you believe?”
This is not what anyone expected to happen first.
Line editing—the supposedly mechanical task—turns out to be harder. AI can fix obvious errors, but it struggles with voice. It over-smooths. It reaches for the expected word instead of the precise one. It can’t tell when a sentence fragment is a mistake versus a deliberate choice. It doesn’t know when breaking a rule serves the piece.
The reason is counterintuitive: developmental editing is more legible. Structure can be described in terms of function. “This section establishes the problem.” “This paragraph provides evidence.” “This transition connects A to B.” These are relationships that can be articulated and therefore learned.
Line editing requires a different kind of knowledge—embodied, intuitive, accumulated through thousands of hours of reading. You know the sentence is wrong before you can say why. The rhythm is off. The word has the wrong temperature. This kind of knowledge doesn’t decompose into rules.
The Judgment Problem
But here’s what remains genuinely hard: editorial judgment.
An AI can tell you whether an argument is internally coherent. It cannot tell you whether the argument matters. It can identify what a piece is saying. It cannot tell you whether that thing is worth saying right now, to this audience, in this publication.
Editorial judgment requires a model of the reader that includes their existing beliefs, their attention constraints, their tolerance for difficulty, and their reasons for reading. It requires a model of the publication—not just its stated mission but its accumulated identity, the implicit promises it has made to its audience.
Most importantly, editorial judgment requires the ability to be surprised. A good editor reads a pitch and thinks: “I wouldn’t have thought of that, but now that you’ve said it, I need to know if it’s true.” This is different from evaluating whether an argument is well-constructed. It’s evaluating whether an argument is interesting—which requires having expectations that can be violated.
AI systems are getting better at simulating surprise. They are not yet capable of genuine surprise, because genuine surprise requires having stakes. An editor who publishes a bad piece loses credibility. An editor who misses a great piece loses opportunity. These losses accumulate into judgment. They are not yet losses an AI can experience.
The New Bottleneck
For decades, the bottleneck in publishing was production. There weren’t enough skilled people to do the cognitive work of editing at scale. Good developmental editors were rare. Good line editors were rare. Publications rationed their attention.
AI has not eliminated these tasks, but it has changed who can do them. A writer with AI assistance can now get structural feedback that used to require an experienced editor. The feedback isn’t as good, but it’s available at 2 AM, for free, without social cost.
The new bottleneck is editorial judgment at scale. We can now produce more competent writing than ever before. The constraint is figuring out which competent writing deserves an audience.
This is not a technical problem. It’s a taste problem. And taste is accumulated through years of reading, arguing, being wrong, and paying the price for being wrong.
What This Requires
If you’re a writer: the ability to self-edit at the structural level is now table stakes. AI handles this. What matters is having something to say that an AI wouldn’t think to say—which means having experiences, positions, and knowledge that aren’t already in the training data.
If you’re an editor: your value has shifted from execution to selection. The question is no longer “can you make this piece better?” The question is “should this piece exist?” That judgment cannot be automated because it requires caring about the outcome.
If you’re building a publication: the moat is taste. Anyone can produce competent content now. The scarce resource is the curatorial voice that tells readers: this is worth your time, and that isn’t. Trust me.
The editor’s eye is the last to go because it’s the only part that requires something at stake.